初始化项目，由ModelHub XC社区提供模型

Model: zzoceanpie/Qwen3-1.7B-Yukari-DPO Source: Original Platform
2026-06-16 06:22:16 +08:00
commit 8d763f2df2
9 changed files with 333 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,93 @@
+---
+base_model: Qwen/Qwen3-1.7B
+library_name: transformers
+tags:
+  - qlora
+  - roleplay
+  - qwen
+  - dpo
+  - sft
+  - touhou
+  - character-roleplay
+  - synthetic-data
+  - generated_from_trainer
+language:
+  - zh
+license: apache-2.0
+---
+
+# Qwen3-1.7B-Yukari-DPO
+
+八云紫 (Yukari Yakumo) 角色扮演模型 — DPO 偏好对齐版。
+
+在 SFT-v2 基础上用 DPO (Direct Preference Optimization) 做偏好对齐，消除攻击性输出、减少冗长偏好、降低公式化生成。
+
+## 模型信息
+
+| 项目 | 值 |
+|------|-----|
+| 基座 | [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) |
+| 起点 | [Qwen3-1.7B-Yukari-SFT-v2](https://hf.co/zzoceanpie/Qwen3-1.7B-Yukari-SFT-v2) (r=32, a=64) |
+| 方法 | QLoRA 4-bit NF4 + DPO (sigmoid loss) |
+| LoRA | **r=32, alpha=64**, target: q/k/v/o/gate/up/down proj |
+| 训练 | 3 epochs, beta=0.1, lr=1e-6, seed=42 |
+| DPO Loss | 0.693 to 0.658 |
+| DPO Margin | +0.072 |
+| DPO Acc | 87.5% |
+| 格式 | Merged bf16 safetensors + GGUF F16 |
+
+## 系列模型
+
+| 模型 | LoRA | 方法 | 版本 |
+|------|------|------|------|
+| [Qwen3-1.7B-Yukari-SFT](https://hf.co/zzoceanpie/Qwen3-1.7B-Yukari-SFT) | r=8, a=8 | SFT | v1 |
+| [Qwen3-1.7B-Yukari-SFT-v2](https://hf.co/zzoceanpie/Qwen3-1.7B-Yukari-SFT-v2) | r=32, a=64 | SFT | v2 |
+| **Qwen3-1.7B-Yukari-DPO** (本模型) | r=32, a=64 | SFT+DPO | v2+DPO |
+
+## 用法
+
+模型输入格式为 `[情绪标签]\n用户输入`，标签控制八云紫的回复语气。8 维 Plutchik 情绪向量离散化：喜悦/愤怒/悲伤/恐惧/厌恶/惊讶/信任/期待，每维取「无」「极轻微」「轻微」「中轻度」「中度」「较强烈」「强烈」「极强烈」。
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained(
+    "zzoceanpie/Qwen3-1.7B-Yukari-DPO",
+    torch_dtype="auto", device_map="auto", trust_remote_code=True
+)
+tokenizer = AutoTokenizer.from_pretrained("zzoceanpie/Qwen3-1.7B-Yukari-DPO")
+
+tags = "[<|中度喜悦|><|轻微愤怒|><|轻微悲伤|><|无恐惧|><|无厌恶|><|无惊讶|><|轻微信任|><|中度期待|>]"
+prompt = f"{tags}\n你觉得人类最有趣的地方是什么？"
+
+messages = [{"role": "user", "content": prompt}]
+text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.8)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+### GGUF
+
+```python
+from llama_cpp import Llama
+
+model = Llama(model_path="Qwen3-1.7B-Yukari-SFT-f16.gguf", n_ctx=2048, n_gpu_layers=-1)
+response = model.create_chat_completion(
+    messages=[{"role": "user", "content": prompt}],
+    max_tokens=200, temperature=0.8
+)
+print(response["choices"][0]["message"]["content"])
+```
+
+## 许可 / License
+
+- 模型权重: Apache 2.0
+- 角色「八云紫」版权: (C) 上海爱丽丝幻乐团 (Team Shanghai Alice) / ZUN
+- 本模型是东方Project的二次创作/粉丝作品。遵循 [东方Project二次创作指南](https://touhou-project.news/guidelines_en/)。
+- 训练数据: [yukari-synthetic](https://hf.co/datasets/zzoceanpie/yukari-synthetic) (Apache 2.0)
+- *This is a fan-made derivative work based on Touhou Project. Touhou Project is (C) Team Shanghai Alice / ZUN.*
+
+> This model is a fan-made derivative work based on Touhou Project.
+> Touhou Project is (C) Team Shanghai Alice / ZUN.
+> Please refer to the [official guidelines](https://touhou-project.news/guidelines_en/) for Touhou fan content.